AI designs bacteria-killing viruses in lab, raising biosecurity concerns

The experiment represents a landmark moment in AI-assisted biology: for the first time, a generative model has designed complete, functional viral genomes that work in real laboratory conditions. The team trained Evo, an AI language model similar to ChatGPT, on approximately 2 million bacteriophages to teach it the rules of viral assembly and gene order. When the model generated 302 novel genomes, 16 proved viable—bacteria cultures showed visible clearings where the AI-designed viruses had killed the host organisms, confirming that the AI had captured real biological constraints rather than merely producing plausible-sounding but inert sequences.

J. Craig Venter, who participated in the work, framed the achievement as "just a faster version of trial-and-error experiments"—the appeal lies in speed. Generative AI can propose and evaluate millions of genetic variants in silico before any physical synthesis, potentially compressed years of drug and biotechnology development timelines. The same logic could apply to fighting bacterial infections in agriculture or designing new gene therapies. Yet Venter's own warnings reveal why this advance unsettles the biosecurity community: the same capability that accelerates beneficial research can, in principle, accelerate harm. He explicitly urged "extreme caution" around viral enhancement work and stated that applying this technique to human pathogens like smallpox or anthrax would trigger "grave concerns." The team's safeguard—excluding human-infecting viruses from the training data—is a procedural firewall but not a technical one. Venter and other experts note that the complexity ceiling is real: extending this approach to more intricate organisms like bacteria would face what biologist Jef Boeke called exponential combinatorial difficulty, "way way more than the number of subatomic particles in the universe." That asymmetry—easy for simple phages, nearly impossible for complex pathogens—offers some reassurance but does not resolve the fundamental question of how such a powerful tool should be governed.